Research On Classification And Indexing In Design Patent Image Retrieval

Design patent images which present the main meaning are the most important partof this patent. Therefore, image content-based design patent retrieval is of greatsignificance. The basic idea of content-based image retrieval is to represent images bythe low-level features of them. However, low-level features often can not describe oneimage accurately and this leads to bad retrieval results. Different to the complexity ofgeneral images in Internet, design patent images have good properties of single object,normalized perspective, clear category and litter blur, etc. Therefore, classification withsupervision could well classify design patent images and narrow the semantic gap.For the shortage of direct retrieval, and the good properties of design patent images,classification model is introduced into the image content-based design patent retrievalsystem to improve the retrieval results. Different to direct retrieval, the query image isfirst classified into a certain visual category by Support Vector Machine (SVM); then,similar images are retrieved in the category. As the classification accuracy could not be100%, retrieval precision of images which are wrongly classified will be very low. Torelieve this problem, similar images are retrieved in the whole database while the imageis probably wrongly classified. Locality Sensitive Hashing (LSH) indexing algorithm isintroduced to ensure the response speed. As the algorithm consumes so much memorythat one single machine could not store all the hash tables in the memory, multi-machineindexing is introduced to relieve the problem.Experiments show that the retrieval precision of this proposed architecture is nearlytwice as high as that of direct retrieval. Precision and recall are both over80%and areover10percent higher than foreign similar systems. The retrieval time is restricted in2seconds. And the memory consumption of each indexing node is affordable.